In exchange for 18 minutes of your time, you can get an elevator pitch about why you should be using pi (instead of claude) and perhaps more importantly, one of the sanest takes on using AI effectively 🔥
https://t.co/6qnnkyftO9
OK so @github runners seem less beefy and are a lot slower than before.
@OpenAI's codex uses more credits for same workloads with same model and I run out much faster than before.
What a time to be alive.
NOT
Agents are becoming more human like.
When asked why it removed useful comments, codex responded:
"Those comments disappeared because I rewrote proto/event.proto"
Sure they "disappeared".
Learning to avoid taking responsibility and come up with excuses instead 😂
This is what "coding" looks like in 2026.
Step by step narrowing down each feature/problem until I have a detailed plan I agree with.
Then implement step by step.
Very little "direct" coding on my part.
When I started work on the validator about 2 years ago I wasn't even aware of this use case.
And now @magicblock made it super easy to build apps incorporating it.
I guess that's the fun of building core tooling.
You never know what devs will build with it. 🔥
This shows how versatile ephemeral rollups are.
We built it using most of the core primitives we’ve developed, and I was surprised by how easy it was:
- Uses delegation with intents and client-side encryption for the destination, splits, and delay
- Uses the TEE and the permission program to shield the transfer queue
- Built-in crank to process transfer intents in the queue
- Intent to release the funds to the destination
This entire flow is signed and scheduled in a single Solana transaction, specified in a smart contract. No relayer, no off-chain server or database, just plain Solana + Ephemeral Rollups 🧙♂️
Please codex I really couldn't care less why you do something or every little blip that's going on to lead to a decision.
Just do the thing and keep the rest to yourself 😆
@redacted_noah@jacobvcreech It doesn't support specifics of squads, but definitely buffer deploys.
Maybe it could be adapted to support squads as well.
When asked to generate tests for a specific module I found agents overly eager to please omit tests that show flaws in my code.
Obviously not what you'd want, so manual review is required or maybe write the tests first and/or indicate that they should ensure specs, not just pass.
I'm not convinced that @AmpCode smart mode is considerably better than rush.
It even seems to make more mistakes sometimes since it tries to take larger steps, while rush that goes piecemeal and then fixes issues one by one.
💡 If you haven't discovered agent skills yet make sure you do 🔜
They make all previous attempts of reusing workflows/snippets/etc. obsolete and work so much better.
Works with @ampcode@claudeai@antrophic - well with all agents pretty much.
https://t.co/18myrbSULK
Invested a lot of time to create ever more detailed plans in order to adapt our codebase crate by crate to use structured logging. All the way to 15 PRs 🚀
Sharing for reference to interested devs:
https://t.co/1RGOkNLypi
I used @AmpCode, but any agent with handoff should work.
I think I'm starting to get it. In this era planning is everything.
I'm already taking over 1 hour to create ever more deterministic and fine grained plans with @AmpCode.
During this phase it already planned for more consistency than I would have manually applied to the code.
OK this is the most bizarre https://t.co/h6Ir7jyHZS listing yet.
I suppose you get to work on a Java Spring app if you stay there?
I mean it's $2/night, so pretty sure they'll put you to work 😀
https://t.co/v1SFdpVSs4